Open source · MIT · v0.3.0

Skills that make
your coding agent think
like a product data scientist.

They decide if a question deserves an analysis, pick the method, and check the result before shipping.

Read the method map→
Theory layer the ceiling: what each decision taught, and the prior it left writing-readouts Measurement framework the floor: the definition of success, and whether the number can be trusted defining-metrics Routing routing-questions one lookup, three questions, one gate Exploratory analytics sizing-opportunities the question, localised and sized can you randomise? Experimentation designing-experiments reading-experiments Causal inference choosing-causal-designs Statistical modeling automating-decisions the decision, made at volume yes no description hypothesis

Scroll sideways

All skills

routing-questions

Routing

One lookup, three questions, one gate. Decides whether this becomes work at all.

defining-metrics

The floor

A metric turned into a computation, a source of truth, a registry entry, and how it will be gamed.

sizing-opportunities

Description

The floor first, then mix or rate, the step, the segment. A hypothesis with a size, never an effect.

designing-experiments

Causation

The four choices that cannot be repaired after launch. MDE comes from the prior store.

reading-experiments

Causation

SRM, exposure, the sequential bound, CUPED, shrinkage. Whether the result is a result.

choosing-causal-designs

Causation

ITS, DiD, synthetic control, matching, IV, and the exit that says there is no control group.

automating-decisions

Prediction

A decision made at volume. Validated out of time, against the base rate, then a holdout for the impact.

writing-readouts

The ceiling

The decision rule first, the result last. The belief, not just the number.

See it route

routing-questions

> did the new onboarding lift D7 retention?

lookuptheory layer2 prior tests on d7_retention, none on onboarding
gatetrust the metric?d7_retention, registered, one source of truth
q1description or change?a change
q2once or continuously?once, a ship-or-kill call
q3who assigned it?you: a 50/50 flag, live since 04‑08

-> reading-experiments

srm · exposure log · sequential bound · cuped · shrinkage

> how much does a paywall test usually move trial starts?

lookuptheory layertrial_starts: 6 entries, median +0.9pp, none expired

-> settled

the store already answers it. no query, no traffic, no ticket.

> is activation down since the pricing change?

lookuptheory layeractivation: no entry, no prior
gatetrust the metric?two dashboards disagree, no owner, no registry entry

-> defining-metrics

the measurement work is the work. not in the meantime.

Run it on your own question

claude
> /plugin marketplace add 0trm/gallop
> /plugin install gallop@gallop

Then ask your agent the question you were about to write a query for. The plugin is the Claude Code route; any agent that reads skills can take the directories instead.